I believe that the problem is psychological. Many years ago, back in the 1990s, wavelet based algorithms were able to outperform humans on detecting tumors. T

I believe that the problem is psychological. Many years ago, back in the 1990s, wavelet based algorithms were able to outperform humans on detecting tumors. The thing is that the algorithms were better on the easy parts of the mammogram, and worse on the hard parts. So researchers thought that humans plus software should do better still, because the humans would focus on the hard parts that they did better and the software would catch the rest. Unfortunately according to a talk that I was at, it didn't work that way. It turns out that radiologists already spend most of their time on the hard parts. So they quickly dismissed the mistakes of theirs that they found as careless errors, and focused on the mistakes of the algorithm in the hard part as evidence that the algorithm didn't work well. And the result was that the radiologists were so resistant to working with the software that it never got deployed. For the same psychological reason I expect radiologists to never voluntarily adopt AI. And they will resist until we reach a point that the decision is taken out of their hands because hospitals face malpractice suits for not having used superior AI.